{"id":"80a37fd4-4149-4da0-af4b-0b19fb76049d","arxiv_id":"2412.06848","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A review of RMT in high-dimensional statistics that contributes a new CLT for the log-eigenvalues of Wishart matrices, with a flawed proof.","lead":"This paper is a review of random matrix theory tools used in high-dimensional statistics, with one new theorem about the log-eigenvalues of Wishart matrices. The theorem is a central limit theorem with an explicit error rate, but the proof has a gap and the manuscript has many typographical errors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof drops the scale factor sqrt(n/(n-p+i)); for p/n -> gamma>0 the standard-normal limit is false, so the theorem needs an explicit p/√n -> 0 restriction.","rationale":"The reader's weakest assumption correctly identifies the scaling mismatch in the proof of Theorem 1. My stress-test sharpens this to show that the theorem is not merely missing a proof step but is false as stated when p/n -> gamma>0. The proof's application of Lemma 3 without the scale factor sqrt(n/(n-p+i)) is the concrete leak, and the resulting variance calculation confirms the limiting distribution is not standard normal in that regime. Since the applications in Section 4.1 already work under p ∼ n^{1/2-ε}, the central claim is salvageable with an explicit regime restriction and a corrected proof. Thus the reader's CONDITIONAL verdict remains appropriate; no adjustment is needed beyond holding the authors to the stated fix.","tokens_in":27730,"tokens_out":10708,"duration_ms":99295,"concrete_test":"Simulate n=2000, p=1000, Σ=I_p. Draw X ~ W_p(I,n), compute Z = sqrt(n/(2p))(Σ_i log λ_i(X) - Σ_i log(n-p+i)). Repeat 10^4 times, compare the empirical CDF with Φ. If the sup-norm distance stays near sup_x|Φ(x/σ)-Φ(x)| with σ=sqrt(2 ln 2) ≈ 1.177 instead of decaying, the theorem's claim fails for p/n=1/2. Alternatively, compute the same sup distance for p=n^α with α=0.4,0.5,0.6 to confirm it decays only when p/√n -> 0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The appendix proof of Theorem 1 (Section 6) applies Lemma 3 to each term sqrt(n/2) log(U_i/(n-p+i)), but Lemma 3 bounds the distance between sqrt((n-p+i)/2) log(U_i/(n-p+i)) and N(0,1). The omitted factor sqrt(n/(n-p+i)) is not close to 1 unless p/n -> 0. For p/n -> gamma>0, the per-term scaling error is O((p-i)/n), summing to O(p^2/n), which is not O(p/√n) unless p/√n -> 0. More seriously, for p/n -> gamma>0, Var(log(U_i/(n-p+i))) ≈ 2/(n-p+i), so Var(sum_i sqrt(n/(2p)) log(U_i/(n-p+i))) ≈ (1/(2γ))·2 log(1/(1-γ)) = log(1/(1-γ))/γ ≠ 1. Hence the limiting distribution is not standard normal; Theorem 1 as stated (n ≥ p, no regime restriction) is false. The applications in Section 4.1 assume p ∼ n^{1/2-ε}, so p/√n -> 0, where the result may be salvageable, but the theorem statement and proof need correction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a review of random matrix theory (RMT) in high-dimensional statistics, covering the spectral properties of sample covariance matrices and F-type matrices and surveying applications in covariance inference, PCA, signal processing, and changepoint detection. Its claimed original contribution is Theorem 1, a Berry-Esseen-type bound for the log-determinant of a Wishart matrix, with applications to one-sample, regression, and two-sample covariance tests in Section 4.1. The proof of Theorem 1 is given in the appendix.","tokens_in":27958,"tokens_out":11840,"duration_ms":106550,"significance":"If Theorem 1 held as stated, it would provide an explicit O(p/sqrt(n)) rate of convergence for a log-determinant CLT and would support new hypothesis tests for covariance matrices. The paper also compiles several classical RMT results and presents them in a statistical context, which is useful for a review. However, the theorem is false as stated without a p/sqrt(n) -> 0 restriction, and the proof omits a nontrivial scaling factor. The test statistics derived from the theorem in Section 4.1 contain additional centering errors, so the claimed applications are not valid in their current form. The survey component is reasonable, but the original contribution needs substantial correction.","major_comments":[{"comment":"Lemma 3 is applied to the term sqrt(n/2) log(U_i/(n-p+i)), but Lemma 3 controls sqrt((n-p+i)/2) log(U_i/(n-p+i)). The omitted factor sqrt(n/(n-p+i)) is not close to 1 uniformly in i unless p/n -> 0; its deviation is O((p-i)/n), which sums to O(p^2/n). For p/n -> gamma > 0, the variance of the normalized statistic is approximately (n/(2p)) * sum_i Var(log(U_i/(n-p+i))) ~ (1/gamma) log(1/(1-gamma)), which is not 1, so the standard normal limit in Theorem 1 is false for such sequences. The theorem must be restated with an explicit p/sqrt(n) -> 0 restriction (or another correct regime), and the proof must bound the distributional distance between the scaled summands and N(0,1) rather than invoking Lemma 3 directly. In addition, the displayed inequality after the reduction to Y_1+...+Y_p does not follow from Lemma 1 as written: the right-hand side still contains the same p-dimensional probability, so Lemma 1 must be applied successively to the individual summands, with each individual term stated explicitly.","section":"Section 6, Appendix (proof of Theorem 1)"},{"comment":"The test in Eq. (4.4) is not a valid application of Theorem 1. With S defined as (1/(n-1)) * sum (X_i - bar X)(X_i - bar X)^T and hat lambda_i its eigenvalues, the Wishart matrix is (n-1)S, not S. Theorem 1 applied to (n-1)S gives the statistic sqrt((n-1)/(2p)) [sum_i log(hat lambda_i/lambda_i) + p log(n-1) - sum_i log((n-1)-p+i)]. The expression in Eq. (4.4) omits the term p log(n-1) and uses log(n-p+i) instead of log((n-1)-p+i), leaving a deterministic offset of order sqrt((n-1)/(2p)) [p log(n-1) + log(n/(n-p))], which diverges as p, n -> infinity. Hence Eq. (4.4) does not define an asymptotically size-alpha test under the stated definitions.","section":"Section 4.1, Eq. (4.4)"},{"comment":"The two-sample test in Eq. (4.8) has an analogous centering error. Under H0, (m-1)S_X ~ W_p(Sigma, m-1) and (n-1)S_Y ~ W_p(Sigma, n-1), so the log-ratio of the Wishart determinants contains an additional p log((m-1)/(n-1)) term, and the centering should be sum_i log(((m-1)-p+i)/((n-1)-p+i)). Eq. (4.8) instead uses sum_i log((m-p+i)/(n-p+i)) and omits the p log((m-1)/(n-1)) term. For n/m -> c with c != 1, the omitted term times the normalization sqrt(m/(2p)(1+1/c)) is of order sqrt(mp) log c and does not vanish, so the proposed test is not asymptotically size alpha. The normalization also does not match the variance of the centered log-ratio.","section":"Section 4.1, Eq. (4.8)"}],"minor_comments":[{"comment":"The index k in Eq. (4.6) is undefined, and the dimension index is inconsistent: the notation writes hat lambda_1, ..., hat lambda_p even though the SSE matrix is m x m. The correct centering should be sum_{i=1}^m log((n-r)-m+i) rather than sum log(n-r-k+i).","section":"Section 4.1, Eq. (4.6)"},{"comment":"Proposition 1 refers to T as defined in (2.5), but the definition of T appears in Eq. (4.23); the cross-reference is incorrect.","section":"Section 4.4, Proposition 1"},{"comment":"The Lindeberg-type conditions in Theorem 12 are labeled (3.7) and (3.8), but they appear in Section 4.4; the equation numbering should be corrected.","section":"Section 4.4, Theorem 12"},{"comment":"The sentence 'For gamma = 0 ... the maximum and minimum eigenvalues converge to 1' is imprecise: it should refer to the support of the limiting spectral distribution, or specify that this concerns the sample eigenvalues under a further scaling.","section":"Section 3.1.1"},{"comment":"There are several typographical and formatting errors, including 'eignevalues' in Section 3.1, 'Tthe generalized factor model' in the reference Forni et al., and inconsistent use of p versus m in parts of Section 4.1. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The current version cannot be accepted because the central new theorem is false as stated and the main test statistics in Section 4.1 are not correctly derived from it. The errors appear local and fixable: adding the p/sqrt(n) -> 0 restriction to Theorem 1, correcting the proof to handle the scaling factor, and fixing the centering in Eqs. (4.4) and (4.8) could make the claims defensible. The review portions are a competent survey, but the original contribution requires careful reworking before the paper can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a broad review of random matrix theory in high-dimensional statistics, with one original theorem: a Berry–Esseen-type bound for the log-eigenvalue sum of a Wishart matrix, giving an explicit O(p/√n) error rate. The review parts are genuinely competent. The survey of MP law, Tracy–Widom, F-type matrices, and the applications to covariance testing, PCA, signal processing, and changepoint detection is mostly accurate and well-organized. The derived tests in Section 4.1 are reasonable and correctly use the CLT under the stated p ~ n^{1/2−ε} regime. The soft spot is the proof of Theorem 1, and it is load-bearing. Lemma 3 applies to sqrt((n−p+i)/2) log(U_i/(n−p+i)), but the proof applies Lemma 3 to sqrt(n/2) times the same log. The omitted factor sqrt(n/(n−p+i)) is not close to 1 unless p/n → 0. The stress-test note is right: for p/n → gamma > 0, the variance of the sum is log(1/(1−γ))/γ, not 1, so the standard normal limit fails. Theorem 1 as stated — with only n ≥ p and no regime restriction — is false. The good news is that the fix is simple: state the theorem under p/√n → 0 (or p/n → 0 with a rate), and add a short argument bounding the scaling factor. The applications in Section 4.1 already assume p ~ n^{1/2−ε}, so the practical content survives. The manuscript is also textually sloppy: typos, broken cross-references, and at least one garbled citation (the Forni reference). These need cleanup but do not affect the math. This paper is for readers who want a broad overview of RMT applications in statistics, and for anyone building covariance tests where p is much smaller than √n. The explicit rate could be citable after correction. It deserves peer review: a serious referee should send it back for major revision, not desk reject. The theorem needs to be restated and reproved correctly, and the text cleaned up, but the review content has value and the central idea is salvageable.","headline":"A useful RMT review whose one original theorem has a real scaling error in the proof; as stated, the theorem is false outside the p/√n → 0 regime, but the applications already live in that regime and the fix is straightforward.","tokens_in":756,"tokens_out":709,"would_cite":false,"duration_ms":28133,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","62H10","62H15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves a Berry–Esseen-type bound for the log-determinant of a Wishart matrix and uses it to build tests for covariance matrices in high dimensions.","keywords":["Random Matrix Theory","Empirical Spectral Distribution","Limiting Spectral Distribution","Wishart matrix","Berry-Esseen bound","covariance testing","high-dimensional statistics"],"falsifier":"Simulate the statistic in Theorem 1 for a fixed ratio $p/n$ (say $p/n=1/2$) with large $n$, and compare the empirical CDF to $\\Phi$. If the error does not shrink at rate $p/\\sqrt{n}$, or the normalization is visibly wrong, the theorem's stated rate fails. More directly, inspect the proof's step where $\\sqrt{n/2}$ is factored out of each log term: the omitted factor $\\sqrt{n/(n-p+i)}$ differs from $1$ by a factor of about $\\sqrt{2}$ at $i=1$ when $p/n=1/2$, so Lemma 3 cannot be applied as written; a simulation at that ratio would show whether the CLT itself survives the mismatch.","tokens_in":27515,"feed_emoji":"📊","tokens_out":5533,"duration_ms":48262,"temperature":0.7,"pith_summary":"This is a review of random matrix theory as a toolbox for high-dimensional statistics, and it carries one original result of its own. Theorem 1 states that for a Wishart matrix $X_n \\sim W_p(\\Sigma,n)$, the centered and scaled sum of log-eigenvalue ratios — equivalently the log-determinant of $X_n$ relative to $\\Sigma$ — converges to a standard normal with a Berry–Esseen error of order $p/\\sqrt{n}$. A sympathetic reader should care because that rate turns a limit theorem into a usable approximation: the paper shows how it yields asymptotically size-$\\alpha$ tests for $H_0:\\Sigma=\\Sigma_0$ in one-sample, regression, and two-sample problems, and how it gives an error bound for estimating functions of the population spectrum by sample eigenvalues. The rest of the article surveys the Marcenko–Pastur law, extreme-eigenvalue Tracy–Widom limits, spiked covariance phase transitions, and $F$-matrix results, and applies them to PCA, signal detection, and changepoint detection.","feed_headline":"A Berry-Esseen rate for Wishart log-determinants enables covariance tests","feed_subtitle":"The paper proves a CLT with O(p/√n) error and builds size-α tests for covariance matrices from it.","key_machinery":"The argument runs on the classical determinant factorization of a Wishart matrix: $|X_n| = |\\Sigma| U_1\\cdots U_p$, where $U_i \\sim \\chi^2_{n-p+i}$ are independent (the $i$-th factor carries the degrees of freedom $n-p+i$). This reduces the log-determinant statistic to a sum of independent log-chi-square terms, and the proof assembles a Berry–Esseen bound for each term (Lemma 3, proved from a delta-method bound of Pinelis and Molzon and the classical Berry–Esseen theorem) with a subadditivity lemma for sup-norm CDF distance (Lemma 1). The machinery delivers the $O(p/\\sqrt{n})$ rate by adding $p$ terms each of order $1/\\sqrt{n-p+i}$.","core_discovery":"On the paper's own terms, the central discovery is Theorem 1: if $X_n \\sim W_p(\\Sigma,n)$ with $n \\ge p$, then $\\sup_x |P( \\sqrt{n/(2p)} (\\sum_i \\log(\\lambda_i^{(n)}/\\lambda_i) - \\sum_i \\log(n-p+i)) \\le x) - \\Phi(x)| = O(p/\\sqrt{n})$. The statement is a CLT with explicit rate for the log-determinant of the Wishart matrix: the random part is exactly $(1/2)\\sum_i \\log(U_i/(n-p+i))$ with $U_i$ independent $\\chi^2$ variables, so the theorem gives the fluctuation scale $n/(2p)$ and the centering $\\sum_i \\log(n-p+i)$. Section 4.1 converts the theorem into concrete testing procedures: an asymptotically size-$\\alpha$ test of covariance equality in equation (4.4), its version for the sum-of-squares error matrix in high-dimensional regression in equation (4.6), and a two-sample test with an explicit approximate power function in equation (4.8). The paper claims these tests work when $p/\\sqrt{n} \\to 0$.","pith_inferences":["The theorem's regime $p/\\sqrt{n} \\to 0$ is much more restrictive than the usual $p/n \\to \\gamma$ regime of random matrix theory: $p$ must be asymptotically smaller than $\\sqrt{n}$, so the tests are not directly usable in the standard proportional-growth setting.","If the rate $O(p/\\sqrt{n})$ is sharp, the normal approximation should degrade once $p$ is comparable to $\\sqrt{n}$; a natural extension would be a non-normal limiting law in which the scaling factor $\\sqrt{n/(n-p+i)}$ is kept inside each log term.","The same machinery could be applied to other smooth functions of Wishart eigenvalues besides the logarithm, such as $\\log\\det(I+S)$ in signal-processing capacity computations, giving explicit convergence rates for those statistics."],"forward_implications":["The statistic in Theorem 1 gives an asymptotically size-$\\alpha$ test of $H_0:\\Sigma=\\Sigma_0$ with critical values from the standard normal (equation 4.4).","The same theorem supplies a test for the error covariance matrix in high-dimensional linear regression using the SSE matrix, which is Wishart under the null (equation 4.6).","For two independent samples, the theorem yields an asymptotic test of $\\Sigma_1=\\Sigma_2$ and an explicit formula for its power function in terms of the population eigenvalues (equation 4.8).","The $O(p/\\sqrt{n})$ rate provides a bound on the error when functions of the population spectrum are approximated by sample eigenvalues, which the paper notes is useful for sample-size determination."],"supporting_citations":[{"why":"Provides the delta-method Berry–Esseen bound used in Lemma 2, which is the backbone of the per-term chi-square error.","marker":"Pinelis and Molzon (2016)"},{"why":"Source for the Wishart distribution and the determinant factorization $|X_n|=|\\Sigma|\\prod U_i$ used to reduce the statistic to independent chi-square terms.","marker":"Muirhead (2009)"},{"why":"Defines the Wishart distribution whose determinant factorization Theorem 1 exploits.","marker":"Wishart (1928)"},{"why":"Supplies the distributional facts used in Section 4.1: the SSE matrix in regression is Wishart under the null, so Theorem 1 applies to the regression test.","marker":"Mardia et al. (2024)"}],"fun_headline_variants":["Wishart log-det CLT with explicit rate enables covariance tests","Random matrix theory yields precise CLT for Wishart log-determinants","Explicit Berry-Esseen rate for Wishart log-det powers testing","New CLT for Wishart determinants drives covariance inference","O(p/√n) error in Wishart log-det CLT builds valid tests"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof treats each term $\\sqrt{n/2}\\,\\log(U_i/(n-p+i))$ as if the chi-square variable had $n$ degrees of freedom rather than $n-p+i$, an identification that is accurate only when $p$ is small relative to $n$.","fun_headline_variants_meta":{"raw":{"variants":["Wishart log-det CLT with explicit rate enables covariance tests","Random matrix theory yields precise CLT for Wishart log-determinants","Explicit Berry-Esseen rate for Wishart log-det powers testing","New CLT for Wishart determinants drives covariance inference","O(p/√n) error in Wishart log-det CLT builds valid tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1593,"prompt_tokens":894,"completion_tokens":699,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":605}},"tokens_in":510,"tokens_out":699,"duration_ms":6519,"temperature":1.0,"reasoning_tokens":605,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:20:52.832064+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the statistic in Theorem 1 for a fixed ratio $p/n$ (say $p/n=1/2$) with large $n$, and compare the empirical CDF to $\\Phi$. If the error does not shrink at rate $p/\\sqrt{n}$, or the normalization is visibly wrong, the theorem's stated rate fails. More directly, inspect the proof's step where $\\sqrt{n/2}$ is factored out of each log term: the omitted factor $\\sqrt{n/(n-p+i)}$ differs from $1$ by a factor of about $\\sqrt{2}$ at $i=1$ when $p/n=1/2$, so Lemma 3 cannot be applied as written; a simulation at that ratio would show whether the CLT itself survives the mismatch.","supporting_citations":[{"cited_title":"and Molzon, R","cited_arxiv_id":null,"evidence_quote":"Provides the delta-method Berry–Esseen bound used in Lemma 2, which is the backbone of the per-term chi-square error."},{"cited_title":"V., Kent, J","cited_arxiv_id":null,"evidence_quote":"Supplies the distributional facts used in Section 4.1: the SSE matrix in regression is Wishart under the null, so Theorem 1 applies to the regression test."}],"review_version":1}